Multi-Objective Physics-Informed Neural Network for Energy-Aware Trajectory Planning of an Industrial Six-DOF Manipulator in Vision-Guided Sorting Tasks

Abstract

The trajectory planning for the industrial manipulators using vision-guided pick-and-place tasks requires simultaneous optimization of cycle time, energy consumptions, and motion smoothness, while respecting the kinematic and the dynamic constraints. Whereas in the conventional polynomial methods produce smooth motion but ignores dynamics, and the data-driven methods capture nonlinearities but may violate the physics. The existing physics-informed neural networks (PINNs) for robotics target inverse dynamics rather than full trajectory generations. This paper proposes IRB-PINN, a multi-objective PINN that jointly enforces six physics-based constraints—Newton-Euler dynamics, energy minimisation, jerk smoothness, joint limits, manipulability, and data fidelity—with adaptive gradient-balanced weighting. The framework is trained on 10,000 simulated trajectories of an ABB IRB 1410 manipulator handling nine object classes across a dual-pallet workstation, and benchmarked against six learning baselines (ANN, SVR, GPR, LSTM, RF, GBT). On the held-out test set, IRB-PINN achieves R² = 0.961 (33.9% RMSE reduction over the strongest baseline), reduces cycle time by 18.4%, energy by 14.7%, and jerk by 23.2%, with statistical significance p < 0.001 (Wilcoxon, Bonferroni-corrected). An ablation study isolates the contribution of each loss term, identifying the dynamics loss as dominant (RMSE penalty 0.114 when removed). On the physical IRB 1410 platform, 500 trials yield Pearson r = 0.89 between simulated and measured cycle times. Inference latency of 0.5 ms per query is compatible with real-time IRC5 controller integration.